On macOS, the formatter is printing signed values as
unsigned, it seems, and the tests are expecting correctly
signed values. These tests were added in
https://github.com/llvm/llvm-project/pull/78609
The runtime implements CONVERT="SWAP", but semantics doesn't like it.
Add it to the relevant table.
Fixes llvm-test-suite/Fortran/gfortran/regression/record_marker_1.f90,
.../unf_io_convert_1.f90, .../unf_io_convert_2.f90, and
.../unf_io_convert_3.f90.
We support specific intrinsic calls like `AMAX0(1.0,2)` that have
heterogeneous argument types as an optional extension in cases where the
specific intrinsic has a related generic intrinsic function capable of
handling the argument types. This feature can't be allowed to apply to
calls where the result of the related generic intrinsic function is not
convertible to the type of the specific intrinsic, as in `AMAX0('a',
'b')`.
Fixes https://github.com/llvm/llvm-project/issues/78932.
CheckIPOSupported is used to test for working LTO since #74520. However, before CMake 3.24 this will test the default linker and ignore options such as LLVM_ENABLE_LLD. As a result, CMake would test whether LTO works with the default linker but builds with another one. In a typical scenario, libomptarget is compiled with the in-tree Clang, but linked with ld.gold, which requires the LLVMgold plugin, when it actually would work with the lld linker (or also fail because the system lld is too old to understand opaque pointers). Using gcc as the compiler would pass the test, but fail when linking with lld since does not understand gcc's LTO format.
Disable LTO by default for now since automatic detection causes too many problems. It causes the openmp-offload-cuda-project buildbot (https://lab.llvm.org/staging/#/builders/151) to fail and LLVM_ENABLE_RUNTIMES=openmp builds will have it implicitly disabled in the vast majority of system configurations anyway.
https://github.com/llvm/llvm-project/issues/78927 contains a case of
fixed-form source in which a Hollerith literal is mistakenly tokenized,
leading to grief later due to apparently unbalanced parentheses.
The source looks like "REAL*8 R8HEAP(SCRSIZE)" and the Hollerith literal
is misrecognized as such because it follows "8R". In order to properly
tokenize Hollerith literals in old comma-free FORMAT statements like "1
FORMAT(3I5HFLANG)", the tokenizer in the prescanner treats a letter
after an integer token ("3I") as a special case. The fix is to do this
only when the characters involved are nested in parentheses and
Hollerith is a possibility.
Fixes https://github.com/llvm/llvm-project/issues/78927.
Some Fortran codes use line continuation as a form of token pasting; see
https://github.com/llvm/llvm-project/issues/78797. This works in
compilers that run a C-like preprocessor and then apply line
continuation to its output; f18 implements line continuation during
tokenization and preprocessing, but can still handle this case.
In the rare case when an identifier is split across two or more
continuation lines, this patch allows its parts to be distinct
preprocessing tokens for the purpose of macro replacemnt. They (or their
replacement texts) can be effectively rejoined later as a single
identifier when the cooked character stream is tokenized in parsing.
Fixes https://github.com/llvm/llvm-project/issues/78797.
The utility function GetRelevantObject() seems to be just wrong for
definability checks for the "base object" of a designator, and that's
all for which it is (now?) used. This leads to some false error messages
in Whizard when data-refs with multiple pointer components are defined.
Simplify, and add more test cases.
Encode BRANCHENTRY bits as bitmask for deduplicated entries.
Reduces BAT section size:
- large binary: to 11834216 bytes (0.31x original),
- medium binary: to 1565584 bytes (0.26x original),
- small binary: to 336 bytes (0.23x original).
Test Plan: Updated bolt/test/X86/bolt-address-translation.test
When an attribute specification statement follows a declaration that
applies the SAVE attribute to a symbol, name resolution can produce a
spurious error message about redundant SAVEs.
Fixes llvm-test-suite/Fortran/gfortran/regression/proc_ptr_9.f90.
In instcombine, when we sink an instruction into a successor block, we try
to clone and salvage all the variable assignments that use that Value. This
is a behaviour that's (IMO) flawed, but there are important use cases where
we want to avoid regressions, thus we're implementing this for the
non-instruction debug-info representation.
This patch refactors the dbg.value sinking code into it's own function, and
installs a parallel implementation for DPValues, the non-instruction
debug-info container. This is mostly identical to the dbg.value
implementation, except that we don't have an easy-to-access ordering
between DPValues, and have to jump through extra hoops to establish one in
the (rare) cases where that ordering is required.
The test added represents a common use-case in LLVM where these behaviours
are important: a loop has been completely optimised away, leaving several
dbg.values in a row referring to an instruction that's going to sink. The
dbg.values should sink in both dbg.value and RemoveDIs mode, and
additionally only the last assignment should sink.
The new accurate algorithm for real MOD and MODULO in the runtime is not
as fast as std::fmod(), which is also accurate. So use std::fmod() for
those floating-point types that it supports.
Fixes https://github.com/llvm/llvm-project/issues/78641.
An || operator in namelist substring bounds checking really needs to be
an && operator so that the substring is viewed as correct only when both
its bounds are valid.
Fixes llvm-test-suite/Fortran/gfortran/regression/namelist_40.f90.
Added an --comp-dir-override option that overrides DW_AT_comp_dir in the
unit die. This allows for llvm-bolt to be invoked from any category and
still find .dwo files.
Fortran allows the name of a generic interface to be the same as the
name of a derived type or specific procedure. When this happens, it
causes the code in module file generation to miss the symbol of a
derived type when scanning for symbols in initialization expressions
that need to be imported. Fix.
The standard requires a compiler to diagnose an incorrect use of a
pointer in a DEALLOCATE statement. The pointer must be associated with
an entire object that was allocated as a pointer (not allocatable) by an
ALLOCATE statement.
Implement by appending a validation footer to pointer allocations. This
is an extra allocated word that encodes the base address of the
allocation. If it is not found after the data payload when the pointer
is deallocated, signal an error. There is a chance of a false positive
result, but that should be vanishingly unlikely.
This change requires all pointer allocations (not allocatables) to take
place in the runtime in PointerAllocate(), which might be slower in
cases that could otherwise be handled with a native memory allocation
operation. I believe that memory allocation of pointers is less common
than with allocatables, which are not affected. If this turns out to
become a performance problem, we can inline the creation and
initialization of the footer word.
Fixes https://github.com/llvm/llvm-project/issues/78391.
An Ew.0(Ee) or Ew.0(Ee) output edit descriptor with a scale factor of
zero is an error condition, unless the output edit descriptor was
originally Gw.0(Ee), list-directed, or just E0/D0 with no .d part.
Fixes https://github.com/llvm/llvm-project/issues/78390.
The argument to the PRESENT() intrinsic function must be the name of a a
whole OPTIONAL dummy argument.
Fixes llvm-test-suite/Fortran/gfortran/regression/present_1.f90.
It's beneficial to have uniform reporting in both `infer-stale-profile`
on and off cases, primarily for logging purposes.
Without this change, BOLT would report "input" staleness in
`infer-stale-profile=0` case (without matching), and "output" staleness
in `infer-stale-profile=1` case (after matching).
This change makes BOLT report "input" staleness in both cases. "Output"
staleness information is printed separately with "BOLT-INFO: inferred
profile..."
The general implementation of intrinsic assignment of derived types in
the runtime support library has a doubly-nested loop: an outer loop that
traverses the components and inner loops than traverse the array
elements. It's done this way to amortize the per-component overhead.
However, this turns out to be wrong when the program cares about the
order in which defined assignment subroutines are called; the Fortran
standard allows less latitude here than we need to invert the ordering
in this way when any component is itself an array. So invert the two
loops: traverse the array elements, and for each element, traverse its
components.
Summary:
We support `--target=nvptx64-nvidia-cuda` as a way to target the NVPTX
architecture from standard CPU. This patch simply uses the existing
support for handling `--offload-arch=native` to also apply to the
standalone toolchain.
There's a few restrictions in the standard on the Fortran pointer
argument (FPTR=) to the intrinsic subroutine C_F_POINTER() that almost
no compilers enforce. Enforce them here with warnings.
Support
R_RISCV_TLSDESC_HI20/R_RISCV_TLSDESC_LOAD_LO12/R_RISCV_TLSDESC_ADD_LO12/R_RISCV_TLSDESC_CALL.
LOAD_LO12/ADD_LO12/CALL relocations reference a label at the HI20
location, which requires special handling. We save the value of HI20 to
be reused. Two interleaved TLSDESC code sequences, which compilers do
not generate, are unsupported.
For -no-pie/-pie links, TLSDESC to initial-exec or local-exec
optimizations are eligible. Implement the relevant hooks
(R_RELAX_TLS_GD_TO_LE, R_RELAX_TLS_GD_TO_IE): the first two instructions
are converted to NOP while the latter two are converted to a GOT load or
a lui+addi.
The first two instructions, which would be converted to NOP, are removed
instead in the presence of relaxation. Relaxation is eligible as long as
the R_RISCV_TLSDESC_HI20 relocation has a pairing R_RISCV_RELAX,
regardless of whether the following instructions have a R_RISCV_RELAX.
In addition, for the TLSDESC to LE optimization (`lui a0,<hi20>; addi a0,a0,<lo12>`),
`lui` can be removed (i.e. use the short form) if hi20 is 0.
```
// TLSDESC to LE/IE optimization
.Ltlsdesc_hi2:
auipc a4, %tlsdesc_hi(c) # if relax: remove; otherwise, NOP
load a5, %tlsdesc_load_lo(.Ltlsdesc_hi2)(a4) # if relax: remove; otherwise, NOP
addi a0, a4, %tlsdesc_add_lo(.Ltlsdesc_hi2) # if LE && !hi20 {if relax: remove; otherwise, NOP}
jalr t0, 0(a5), %tlsdesc_call(.Ltlsdesc_hi2)
add a0, a0, tp
```
The implementation carefully ensures that an instruction unrelated to
the current TLSDESC code sequence, if immediately follows a removable
instruction (HI20 or LOAD_LO12 OR (LE-specific) ADD_LO12), is not
converted to NOP.
* `riscv64-tlsdesc.s` is inspired by `i386-tlsdesc-gd.s` (https://reviews.llvm.org/D112582).
* `riscv64-tlsdesc-relax.s` tests linker relaxation.
* `riscv-tlsdesc-gd-mixed.s` is inspired by `x86-64-tlsdesc-gd-mixed.s` (https://reviews.llvm.org/D116900).
Link: https://github.com/riscv-non-isa/riscv-elf-psabi-doc/pull/373
Reviewed By: ilovepi
Pull Request: https://github.com/llvm/llvm-project/pull/79239
Move logic to replace congruent IV increments to helper function, to
reduce the indentation by using early returns. This is in preparation
for a follow-up patch.
We should diff against the base branch, not always against `main`. This
allows the BuildKite pre-commit CI to work properly when we target other
branches, such as `release/18.x`.
We used to support a /branch comment to specify a branch with commits to
backport to the release branch. However, now that we can use pull
requests this is not needed.
This also simplifies the process, because now the cherry-pick job can
create the pull request directly instead of having it split across two
separate jobs.
We recently noticed that the unwrap_iter.h file was pushing macros, but
it was pushing them again instead of popping them at the end of the
file. This led to libc++ basically swallowing any custom definition of
these macros in user code:
#define min HELLO
#include <algorithm>
// min is not HELLO anymore, it's not defined
While investigating this issue, I noticed that our push/pop pragmas were
actually entirely wrong too. Indeed, instead of pushing macros like
`move`, we'd push `move(int, int)` in the pragma, which is not a valid
macro name. As a result, we would not actually push macros like `move`
-- instead we'd simply undefine them. This led to the following code not
working:
#define move HELLO
#include <algorithm>
// move is not HELLO anymore
Fixing the pragma push/pop incantations led to a cascade of issues
because we use identifiers like `move` in a large number of places, and
all of these headers would now need to do the push/pop dance.
This patch fixes all these issues. First, it adds a check that we don't
swallow important names like min, max, move or refresh as explained
above. This is done by augmenting the existing
system_reserved_names.gen.py test to also check that the macros are what
we expect after including each header.
Second, it fixes the push/pop pragmas to work properly and adds missing
pragmas to all the files I could detect a failure in via the newly added
test.
rdar://121365472
Adding the debug hardening modes into PowerPC target to check the
assertion messages, to fix the PPC rehl bot [1]. This is needed as
58780b811c23 changed the assertion to trap in production hardening modes.
Properly fixing these assertions is tracked by #79216.
[1]: https://lab.llvm.org/buildbot/#/builders/57/builds/32354
Co-authored-by: Maryam Moghadas <maryammo@ca.ibm.com>
My previous patch, "Re-exec TSan with no ASLR if memory layout is incompatible on Linux (#78351)" (0784b1eefa36d4acbb0dacd2d18796e26313b6c5) hoisted the 'personality' call, to share the code between Android and non-Android Linux. Unfortunately, this eager call to 'personality' may trigger sandbox violations on non-Android Linux.
This patch fixes the issue by only calling 'personality' on non-Android Linux if the memory mapping is incompatible. This may still cause a sandbox violation, but only if it was going to abort anyway due to an incompatible memory mapping.
(The behavior on Android Linux is unchanged by this patch or the previous patch.)
A quick examination suggests that the current code in the codebase does
not lead to incorrect annotations. However, the intention is for the
object after the function to be annotated in a way that only its
contents are unpoisoned and the rest is poisoned. This commit makes it
explicit and avoids potential issues in future.
In addition, I have implemented a few tests for a function that helped
me identify the specific argument value.
Notice: there is no known scenario where old code results in incorrect
annotation.
40dcf24522af91ab22af2e69f28d1f1d2a860f5c had changed the format
specifier to fix the build for their local system. Unfortunately,
that disagrees with some other systems, such as this buildbot:
https://lab.llvm.org/buildbot/#/builders/37/builds/30440
This patch fixes the issue for all systems by casting.
Extends `vector.insert_strided_slice` and `vector.insert_strided_slice`
to allow scalable input and output vectors. For scalable sizes, the
corresponding slice size has to match the corresponding dimension in the
output/input vector (insert/extract, respectively).
This is supported:
```mlir
vector.extract_strided_slice %1 {
offsets = [0, 3, 0],
sizes = [1, 1, 4],
strides = [1, 1, 1] } : vector<1x4x[4]xi32> to vector<1x1x[4]xi32>
```
This is not supported:
```mlir
vector.extract_strided_slice %1 {
offsets = [0, 3, 0],
sizes = [1, 1, 2],
strides = [1, 1, 1] } : vector<1x4x[4]xi32> to vector<1x1x[2]xi32>
```
The use of SmallSetVector saves 0.58% of heap allocations during the
compilation of a large preprocessed file, namely X86ISelLowering.cpp,
for the X86 target. During the experiment, the final size of ToCheck
was 8 or less 88% of the time.
This patch fixes:
lldb/source/Plugins/Language/CPlusPlus/LibCxx.cpp:1108:39: error:
format specifies type 'long long' but the argument has type
'std::time_t' (aka 'long') [-Werror,-Wformat]
lldb/source/Plugins/Language/CPlusPlus/LibCxx.cpp:1116:64: error:
format specifies type 'long long' but the argument has type
'std::time_t' (aka 'long') [-Werror,-Wformat]
Turns out I was using DbgMarker::getDbgValueRange rather than the helper
utility in Instruction::getDbgValueRange, which checks for null-ness.
Original commit message follows.
[DebugInfo][RemoveDIs] Convert debug-info modes when loading bitcode (#78967)
As part of eliminating debug-intrinsics in LLVM, we'll shortly be
pushing the conversion from "old" dbg.value mode to "new" DPValue mode
out from when the pass manager runs, to when modules are loaded. This
patch adds that conversion process and some (temporary) options to
llvm-lto{,2} to help test it.
Specifically: now whenever we load a bitcode module, consider a flag of
whether to "upgrade" it into the new debug-info mode, and if we're
lazily materializing functions then do that lazily too. Doing this
exposes an error in the IRLinker/materializer handling of DPValues,
where we need to transfer the debug-info format flag correctly, and in
ValueMapper we need to remap the Values that DPValues point at.
I've added some test coverage in the modified tests; these will be
exercised by our llvm-new-debug-iterators buildbot.
This upgrading of debug-info won't be happening for the llvm18 release,
instead we'll turn it on after the branch date, thenbe push the boundary
of where "new" debug-info starts and ends down into the existing
debug-info upgrade path over the course of the next release.
